Carton intelligent right and interest exchange recommendation method and system based on multi-modal data fusion and dynamic portrait

Abstract

The invention relates to the technical field of intelligent recommendation and data fusion, and particularly discloses a carton intelligent right and interest exchange recommendation method and system based on multi-modal data fusion and dynamic portraits. According to the method, physical attributes, user code scanning behaviors and position information of cartons are collected, a heterogeneous graph structure is constructed, and a graph neural network is fused to generate user and carton embedding data; performing collaborative filtering recommendation in combination with historical exchange behaviors of the user, and performing weight correction by using a strategy rule plug-in; further collecting user feedback, updating model parameters through federal learning, and constructing a negative sample set to train a meta learning model to generate a right conversion rule; and generating a life cycle label in combination with a carton circulation track, and dynamically adjusting the right priority. According to the method, the personalized precision and real-time updating capability of right recommendation are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for recommending intelligent carton rights exchange based on multimodal data fusion and dynamic profiling. Background Art

[0002] With the widespread adoption of green logistics, environmentally friendly packaging, and recyclable cartons, equity incentive mechanisms for carton recycling are gaining increasing attention. Existing carton equity redemption recommendation systems are mostly based on static rules or single user behavior analysis models. These systems lack the ability to integrate and process multimodal data, such as carton attributes, user scan frequency, and location information. This makes it difficult to dynamically perceive and accurately model user preferences and the carton lifecycle.

[0003] Traditional recommendation methods, such as collaborative filtering and content recommendations, mostly rely solely on historical click or rating records, failing to fully consider key dynamic data such as changes in the physical characteristics of cartons, user behavior patterns across time and space, and the movement of cartons. Furthermore, existing systems generally lack reverse learning mechanisms for unclicked items, making it impossible to build negative sample inference optimization models, resulting in insufficient recommendation accuracy and real-time responsiveness.

[0004] In this context, there is an urgent need for a multimodal data-driven method that integrates user scanning behavior, carton attribute data and location information, combined with graph neural network embedding modeling and federated learning optimization mechanism, to achieve continuous perception of changes in user rights and interests preferences, dynamic correction of recommendation strategies, and accurate identification of carton life cycle status, thereby improving the personalization, real-time and intelligence level of carton rights and interests recommendations. Summary of the Invention

[0005] The present invention provides a method and system for recommending intelligent carton rights and interests redemption based on multimodal data fusion and dynamic profiling, so as to solve the problem of how to realize personalized and real-time updated intelligent recommendation of carton rights and interests based on the physical attribute data of the carton, the user's scanning behavior and location information, and the integration of graph neural network embedding and dynamic recommendation strategy.

[0006] In order to solve the above technical problems, the present invention provides a carton intelligent rights exchange recommendation method based on multimodal data fusion and dynamic profiling, comprising: Obtain the physical attribute data of the carton, the user's scanning behavior data and location information, perform synchronous filtering and structured coding, and generate standardized feature data; Based on the standardized feature data, a heterogeneous graph structure of user nodes, carton nodes, and equity nodes is constructed, connection strength analysis and pruning are performed, and user embedding data and carton embedding data are generated through a graph neural network; Based on the user embedded data and carton embedded data, combined with the user's historical redemption records, collaborative filtering recommendations are performed, and a strategy rule plug-in is loaded to modify the recommendation weights to generate real-time recommendation list data; Collecting the user's click behavior and redemption results on the real-time recommendation list data, encrypting and compressing it locally, and uploading it to the cloud to perform federated learning training and generate update strategy parameters; Obtain unclicked equity items in the real-time recommendation list data, construct a negative sample set, train a meta-learning model based on the updated policy parameters, generate equity conversion rules, and encapsulate them as a plug-in to be loaded into the policy rule plug-in module; The updating strategy parameter training meta-learning model includes: ; in, For training Meta-learning loss function of ; Recommend function for strategy; Indicates the number of negative samples currently used to train the meta-learning model; is the policy function parameter of the current training; For the The composite feature vector of unclicked equity; It is a candidate replacement equity item; is a set of alternative equity items; Candidate replacement equity Feature embedding of is the square of the Euclidean distance; is the weight coefficient; is a measure of the structural difference between the candidate equity item and the original equity item; Obtain the carton scanning frequency and location migration data reported by the scanning terminal, build a carton flow trajectory record, generate a carton life cycle label, and adjust the equity type priority in the real-time recommendation list data in combination with the policy rule plug-in to generate update policy data.

[0007] Furthermore, generating standardized feature data includes the following steps: Obtain the physical attribute data of the carton, user scanning behavior data and location information; Perform feature extraction and normalization on the physical attribute data of the carton to generate physical attribute features; The user's scanning behavior data and location information are time-series encoded and fused with the physical attribute features to generate standardized feature data.

[0008] Furthermore, generating user embedded data and carton embedded data includes the following steps: Constructing a heterogeneous graph structure including user nodes, carton nodes, and equity nodes based on the standardized feature data; Performing connection strength analysis on the heterogeneous graph structure and performing edge pruning processing; Graph neural network embedding processing is performed on the pruned heterogeneous graph structure to generate user embedding data and carton embedding data.

[0009] Furthermore, generating real-time recommendation list data includes the following steps: Based on the user embedding data and the carton embedding data, combined with the user's historical redemption records, calculating collaborative filtering similarity; The similarity calculation results are input into the policy rule plug-in module, and the rule weights are loaded to perform the recommended weight correction; A preset number of equity items are filtered according to the revised recommendation weights to generate real-time recommendation list data.

[0010] Furthermore, collecting user behavior and performing federated learning training includes the following steps: Collecting the user's click behavior and redemption results on the real-time recommendation list data, and performing encryption and structural compression; Upload encrypted and compressed user click behaviors and redemption results to the cloud, aggregate multi-terminal data, and perform federated learning training; Generate update policy parameters and prepare to send them to edge terminals.

[0011] Furthermore, the method further comprises the following steps: Sending the update policy parameters to the edge terminal; Replace the original model parameters at the edge terminal; Participate in subsequent recommendation weight correction operations based on the updated model parameters.

[0012] Furthermore, generating the equity conversion rules includes the following steps: Obtaining unclicked equity items in the real-time recommendation list data to construct a negative sample set; A meta-learning model is trained based on the negative sample set and the updated strategy parameters to generate new equity conversion rules; The equity conversion rules are encapsulated as a structured plug-in and loaded into the policy rule plug-in module.

[0013] Furthermore, generating a carton life cycle label includes the following steps: Obtain carton scanning frequency and location migration data reported by scanning terminals to build a carton flow trajectory record; Estimate the remaining use cycle of the carton based on its circulation trajectory and generate a carton life cycle label; Output the carton life cycle label for subsequent recommendation strategy use.

[0014] Furthermore, the generating of update policy data comprises the following steps: Combining the carton lifecycle tag with the policy rule plug-in; Adjusting the priority of the equity type in the real-time recommendation list data; Generate updated recommendation strategy data.

[0015] Furthermore, a carton intelligent rights exchange recommendation system based on multimodal data fusion and dynamic profiling includes: Multimodal data acquisition module, used to obtain the physical attribute data of the carton, user scanning behavior data and location information, and generate standardized feature data; A heterogeneous graph construction module, configured to construct a heterogeneous graph structure based on the standardized feature data and perform embedding calculations to generate user embedding data and carton embedding data; A recommendation generation module, configured to perform collaborative filtering calculations based on the embedded data and user history records, and to call a policy rule plug-in to generate real-time recommendation list data; Feedback training module, which is used to collect user feedback, perform federated learning training, and issue updated policy parameters; The rule generation module is used to construct a negative sample set, train the meta-learning model to generate equity conversion rules, and encapsulate them as a plug-in and load them into the policy rule plug-in module; The life cycle identification module is used to generate carton life cycle labels based on carton flow trajectory data, adjust the recommendation strategy priority, and generate update strategy data.

[0016] The key innovations of the present invention include: (1) Heterogeneous graph structure construction and graph neural network embedding fusion mechanism: For the first time, the user, cardboard and equity relationship is modeled as a heterogeneous graph in the cardboard scenario, and the graph structure is optimized through edge pruning and weight analysis, which improves the ability to mine potential correlations between nodes.

[0017] (2) Federated learning-driven policy update path: Building a federated training process based on multi-terminal data feedback aggregation effectively solves the model update challenge under the distributed terminal deployment conditions in the cardboard box scenario.

[0018] (3) Meta-learning reasoning engine and rule plug-in encapsulation mechanism: Use residual samples to construct a negative sample set and train the meta-learning model to generate equity conversion rules with generalization capabilities, and encapsulate them as structural plug-ins for dynamic loading of subsequent recommendation strategies, thereby improving the system's autonomous optimization capabilities.

[0019] (4) Lifecycle identification and dynamic adjustment strategy of rights and interests priority: Introduce scanning frequency and location migration to build a carton flow trajectory record, and adjust the rights and interests recommendation priority in real time based on its lifecycle status, thereby enhancing the system's responsiveness to the carton usage status and recommendation accuracy.

[0020] The following are its main beneficial effects: (1) Significantly improve the accuracy and adaptability of personalized recommendations: This invention constructs a heterogeneous graph structure of user nodes, carton nodes, and equity nodes, and introduces a graph neural network for embedded expression. Compared with the traditional recommendation method based on similarity matching, it can extract deeper structural semantic relationships from multimodal data (including physical attributes, scanning behavior, and location migration), thereby achieving more accurate equity recommendations.

[0021] (2) Strengthening the model's real-time feedback and self-evolution capabilities: The present invention collects users' click behaviors and redemption results, and based on the federated learning mechanism, completes the iterative update of model parameters without uploading the original data, significantly improving the model's strategic adaptability under terminal data privacy protection, and reducing cold start and generalization errors.

[0022] (3) Enhance the dynamic adjustment capability of the equity recommendation mechanism: By introducing the equity conversion rule plug-in constructed by the meta-learning model and combining the carton life cycle label to dynamically adjust the equity priority in the recommendation list, the problem of the static recommendation mechanism's delayed response to changes in user preferences in actual scenarios is effectively avoided, and the "life cycle-behavior feedback-strategy correction" closed-loop optimization is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of a method for recommending intelligent carton benefits exchange based on multimodal data fusion and dynamic profiling provided in an embodiment of the present application; Figure 2 This is a structural block diagram of the carton intelligent rights exchange recommendation system based on multimodal data fusion and dynamic profiling provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] Example 1: Reference Figure 1 , is a flow chart of a method for recommending intelligent carton rights exchange based on multimodal data fusion and dynamic profiling provided by an embodiment of the present invention. The flow may include at least steps S100-S600: Step S100 at least includes steps S110-S130: S110. Obtain physical property data of the carton, user scanning behavior data, and location information through a code scanning terminal, and perform simultaneous filtering and structured coding.

[0025] This step first connects to the front-end system through the code scanning terminal to obtain three types of core data of the carton in real time during the code scanning process: Physical attribute data of the carton, including but not limited to the carton's external dimensions (such as length, width, and height), material parameters (such as corrugated grade), weight grade, and load-bearing parameters; User scanning behavior data, mainly including scanning timestamp, scanning frequency, scanning position offset, scanning force, and other information related to user operation behavior; Location information data, including the latitude and longitude coordinates of the location at the time of scanning, area identification code, warehouse or distribution point number, and other environmental spatial information.

[0026] These three types of data are characterized by time asynchrony, format diversity, and acquisition noise. To achieve unified processing, the system first synchronizes and calibrates the time of these three types of data. It then establishes a data synchronization alignment mechanism based on a unified timestamp window, eliminating redundant information generated by abnormal operations such as device jitter and code scanning failures.

[0027] After synchronization is complete, the system annotates the structured data set with standard fields, fills in missing fields, and corrects illegal fields. Ultimately, it generates a coded data structure with a clear structure, consistent time base, and a fusible format, providing directly callable input data for subsequent feature extraction steps.

[0028] S120 : performing feature extraction and normalization processing on the physical property data of the carton to generate physical property features.

[0029] This step executes a multi-dimensional feature processing flow based on the carton physical attribute data after synchronous encoding in S110. Specifically, it includes: Extract the size parameters and load-bearing parameters of the carton, unify the units, remove outliers, and map the values; Normalize the extracted physical parameters of each type of cartons to a standard dimension range, so that different types of cartons are comparable in the feature space; Based on the preset field structure and graph modeling requirements, a physical attribute feature vector structure with complete semantics is constructed and stored in a standard feature structure.

[0030] The physical attribute features are used as static label information to construct the initial node attributes of the carton node in the subsequent heterogeneous graph construction stage, and are embedded together with user behavior features in the graph neural network to achieve joint modeling of user preferences and scene characteristics.

[0031] S130: Perform time-series coding processing on the user scanning behavior data and the location information, and fuse them with the physical attribute features to generate standardized feature data.

[0032] This step builds a time series vector based on the structured coded user scanning behavior data and location information in S110 and integrates physical attribute features to generate standardized feature data. Specifically, it includes: Encode the user's scanning time series to generate scanning cycle distribution, operation density matrix and behavior pattern description indicators; Map location information into a spatial grid structure, and combine the scan point offset to calculate spatiotemporal behavior indicators such as user activity radius and behavior center offset; The above-mentioned user scanning behavior characteristics and location information characteristics are combined to form a dynamic feature structure; The dynamic feature structure is connected with the physical attribute features to generate complete standardized feature data, and the corresponding carton ID and user ID are marked.

[0033] By synchronously acquiring, structuring, and integrating features from the carton's physical attribute data, user scanning behavior data, and location information in this step, the system achieves efficient conversion of multimodal data into standardized feature structures. This standardized feature data not only provides a unified, high-density, and low-redundancy data input source for heterogeneous graph structure construction and graph neural network embedding, but also lays the data foundation for subsequent recommendation strategy generation and feedback tuning, ensuring that the recommendation system's behavioral responses in dynamic scenarios possess structural rationality, behavior recognition capabilities, and scenario adaptability.

[0034] Step S200 at least includes steps S210-S230: S210: Construct a heterogeneous graph structure of user nodes, carton nodes, and equity nodes based on the standardized feature data.

[0035] Specifically, the system first calls the standardized feature data generated in S130. The standardized feature data includes three parts: user scanning behavior characteristics, location information characteristics, and carton physical attribute characteristics. The system uses the unique identification field to bind this feature data with the user identity, carton unique code, and historical rights and interests association records, and constructs the following three types of nodes: User node: Represented by each user who scans the QR code. Its node attribute fields are composed of scanning behavior characteristics and location information characteristics, including scanning frequency, scanning density distribution, scanning time window code, geographic location information code, etc. Carton node: represented by each carton entity. Its node attribute fields are composed of physical attribute features extracted from S120, including carton size parameters, load-bearing parameters, number of corrugated layers, structure type, etc. Equity nodes: Represented by various types of redeemable equity, the initial node attributes are constructed through the associated equity type label, applicable carton range, and adaptive regional range.

[0036] The system then establishes edges between the user node and the carton node based on the user's code scanning history and the carton code scanning events. It also establishes edges between the carton node and the equity node based on the actual carton redemption records. All edges are bound to attribute fields, which are derived from the raw values ​​and combined features of the standardized feature data and are used to subsequently calculate connection strength and neighbor node embedding weights.

[0037] S220 , performing dynamic connection strength analysis and edge pruning on the heterogeneous graph structure to retain high-weight connection relationships.

[0038] After completing the initial heterogeneous graph construction, the system further performs dynamic connection strength analysis on the connection edges in the graph structure to eliminate low-value edge relationships, reduce computational redundancy, and improve the learning efficiency of the graph neural network.

[0039] Specifically, the system calls various connection edges in the graph constructed in S210 and performs strength analysis based on the following information: For the connection edge between the user node and the carton node, the system analyzes the user's scanning frequency, scanning time density and geographical proximity to the carton, and generates a multi-dimensional connection weight index; For the connection edges between the carton node and the equity node, the system analyzes the historical exchange frequency, the exchange success rate, and the degree of compatibility between the carton and the equity in terms of physical and geographical attributes to generate a comprehensive weight factor.

[0040] After completing the connection strength modeling described above, the system sets a weight threshold and performs edge pruning, retaining connections with weights greater than the threshold and removing weak connections or edges without behavioral support. This pruning operation does not change the original graph's node set, but only acts on the edge set. This improves computational efficiency during the embedding learning phase while preserving the graph's semantic structure.

[0041] S230. Perform graph neural network embedding operations on the pruned heterogeneous graph structure to generate user embedding data and carton embedding data.

[0042] In this step, the system performs graph neural network embedding operations based on the pruned heterogeneous graph structure generated in S220 to extract high-order semantic vector representations of user nodes and carton nodes.

[0043] Specifically, the system uses the corresponding initial node attributes as embedding input for each node category. For user nodes, it uses code scanning behavior and location information encoding features; for carton nodes, it uses physical attribute features; and for equity nodes, it uses equity type and adaptation range labels.

[0044] The system employs a heterogeneous graph neural network architecture to aggregate neighboring node information based on node category and edge type. Each round of graph convolution merges the embedding information of adjacent nodes with the target node based on pruned edges, forming an updated node representation. This process gradually propagates the contextual semantics of the graph structure through aggregation at different levels, ensuring that the final representation of each user node or carton node not only includes its own attribute features but also the semantic information of the entities with which it interacts.

[0045] During the graph neural network training process, the preset graph structure embedding strategy and contrastive learning method are used to update parameters. After the training is completed, the system exports the user embedding data and the carton embedding data and stores them in a structured embedding vector format for use in the collaborative filtering recommendation in the S300 module.

[0046] The user embedded data and the carton embedded data are matched with the user's historical redemption records in S300 and input into the policy rule plug-in module to participate in the recommendation weight calculation.

[0047] By constructing a heterogeneous graph structure, performing connection pruning, and graph neural network embedding processing in this step, the system achieves multi-dimensional semantic joint modeling of user behavior, physical attributes, and equity relationships in the carton scene. User embedded data and carton embedded data have higher-dimensional representation capabilities in subsequent collaborative filtering recommendations, significantly improving recommendation accuracy and real-time matching capabilities. At the same time, graph structure sparsification effectively reduces the training burden of the graph neural network, making the system suitable for running in low-computing environments such as edge devices, and providing a stable and efficient feature extraction and modeling foundation for large-scale carton intelligent equity recommendation systems.

[0048] Step S300 at least includes steps S310-S330: S310: Based on the user embedded data and the carton embedded data, combined with the user's historical redemption records, perform collaborative filtering similarity calculation.

[0049] Specifically, the system calls the user embedding data and carton embedding data generated in module S230 and extracts historical redemption records corresponding to the current user ID to form a three-dimensional input structure for recommendation. The user embedding data is learned by user nodes during the heterogeneous graph neural network embedding process and includes scanning behavior patterns, spatial trajectory preferences, and behavior frequency expressions. The carton embedding data reflects a comprehensive representation of the carton in terms of physical attributes and user behavior interaction.

[0050] In this step, the system performs a similarity matching operation between the current user's embedded data and all carton embedded data. To improve the scenario adaptability of the similarity calculation, the system incorporates the carton identification code, redemption frequency, and post-redemption response feedback information contained in the user's historical redemption records as part of the matching weighting.

[0051] Based on the three-dimensional structure above, the system performs collaborative filtering calculations to generate a set of initial recommendation scores. These scores describe the strength of the current user's match with the equity associated with each candidate carton in the system. These recommendation scores serve as the basis for weighting adjustments in the next step.

[0052] S320: Input the similarity calculation result into the policy rule plug-in module, load the rule weight and perform the recommended weight correction.

[0053] In this step, the system inputs the recommendation score generated in S310 into the policy rule plug-in module. The policy rule plug-in module includes multiple preset rule plug-ins, each of which encapsulates specific rights adaptation logic, user group label mapping rules and real-time scenario priority configuration functions.

[0054] Specifically, the system dynamically loads the corresponding rule plug-in based on the user's current region, the current carton lifecycle status (which is updated and maintained by the S600 module), and the equity tag matching rules in the policy rule plug-in library. Each rule plug-in contains a set of rule weight parameters, which are then weighted and integrated with the input recommendation score.

[0055] For example, for users in a specific region, the plug-in can load a function to adjust the priority of benefits based on geographic preferences. For scenarios where cartons are at the end of their lifecycle, the plug-in can load a strategy to increase the priority of benefits recommended for high recycling value. The system combines these rule functions with the original recommendation score to perform weighted adjustments and generate a recommendation weight vector.

[0056] This recommendation weight vector will be passed to step S330 as the core reference indicator for screening and sorting equity items, forming a dynamically adjustable personalized recommendation result.

[0057] S330: Filter a preset number of equity items according to the revised recommendation weights to generate real-time recommendation list data.

[0058] In this step, the system calls the recommendation weight vector generated in S320 to perform equity item screening and sorting operations.

[0059] Specifically, the system sorts the recommended weight vectors in descending order according to the upper limit of the number of recommended items set by the current business configuration, and selects the top several equity items. The selected equity items must meet the following conditions at the same time: The recommendation weight is higher than the set threshold; The matching degree of rights and interests adaptation carton attribute labels is high; The benefit is valid in the current geographical location and life cycle state; The user has not made the same redemption in history, or the repeated redemption is not within the blacklist restriction period.

[0060] The above screening strategy ensures that the recommendation results are differentiated, effective, and adaptable in real time. The system structures the filtered benefits and generates a recommendation list structure containing benefit identifiers, display copy, redemption instructions, etc., and pushes it to the user interface.

[0061] The final output of real-time recommendation list data is not only used for user interaction display, but will also be passed to the S400 module to collect click behavior and redemption feedback, supporting the federated learning strategy update mechanism.

[0062] Through the S300 module's collaborative filtering recommendations, policy rule plug-in modification, and real-time list generation mechanism, the system implements real-time, personalized benefit recommendations driven by multimodal data in the cardboard box scenario. The semantic similarity space constructed from embedded data ensures the individual matching of recommendations, while the rule plug-in module implements cross-scenario policy mapping and differentiated configuration, making recommendation results adaptable to user operations, cardboard box lifecycle status, and geographic location. This mechanism effectively improves the accuracy and flexibility of recommendation matching, providing a stable and scalable recommendation decision-making foundation for subsequent feedback optimization and policy migration.

[0063] Step S400 at least includes steps S410-S430: S410: Collect the user's click behavior and redemption results on the real-time recommendation list data, and perform local encryption and structural compression.

[0064] Specifically, the system first calls the real-time recommendation list data output in S330, which includes the rights and interests items currently visible to the user, the order in which the rights and interests items are sorted, and the recommendation weight information corresponding to each item.

[0065] When users browse the recommendation list, the system collects the following behavioral data in real time: The timestamp and location where the user clicks on a benefit item; Whether the user has completed the redemption of the equity item; Whether the user has behavioral characteristics such as fast sliding, long pauses, and repeated clicks during the operation; The carton code, redemption rights code and current location information associated with the user's click behavior.

[0066] The above data forms the original user behavior log, which also contains the context information at the time of recommendation, such as the loaded policy rule plug-in identifier and the recommendation weight correction parameter at that time.

[0067] To ensure data privacy and edge device transmission efficiency, the system performs local encryption and structural compression on the original user behavior logs. Specifically, it includes: Use asymmetric keys to desensitize and encrypt the user identification field and the rights identification field; A sliding window structure conversion method is used for the behavior sequence to compress the redundant time series; The unstructured click trajectory is vectorized and mapped and represented using a predefined sparse matrix structure.

[0068] Finally, an encrypted and compressed user feedback data structure is generated, providing data input for uploading to the cloud for joint training.

[0069] S420: Upload the encrypted and compressed user click behavior and redemption results to the cloud, and aggregate multi-terminal data for federated learning training.

[0070] In this step, the system uploads the encrypted compressed structure generated in S410 to the federated learning cloud platform and performs multi-terminal data aggregation.

[0071] Specifically, the system archives and aggregates uploaded data by time window based on the geographical area, scanning device number, and operation node identifier of each edge terminal, and labels it based on the user group behavior dimension. The aggregated data includes: Trends in recommendation success rates across different regions; Click-through conversion rate statistics of each strategy plug-in combination in different terminals; Matching stability between different carton types and equity combinations.

[0072] Based on this multi-source user feedback data, the system loads a pre-configured federated learning training framework in the cloud and initiates global policy model training. During training, feedback data uploaded by each edge terminal is only used for local model gradient calculations; the original data is not stored in the cloud, ensuring data privacy and compliance at the edge.

[0073] The global model training task is based on the following two core goals: Optimize the recommendation weight calculation strategy to make it more in line with user click behavior and redemption preferences; Adjust the adaptation parameters of the policy rule plug-in to improve the scenario adaptability and execution efficiency of rule calls.

[0074] During the training process, the system uses methods such as batch aggregation, parameter relay, and gradient perturbation to ensure that the data from each terminal contributes evenly to the training, and dynamically corrects extreme deviations.

[0075] After training is completed, the system generates updated policy parameters and generates a deployable policy module version for each edge terminal.

[0076] S430: Obtain update strategy parameters after the federated learning training, and send them to the edge terminal to replace the original model parameters.

[0077] In this step, the system packages the updated policy parameters generated after training in S420 into a standard policy model structure, manages them separately according to terminal type and policy version, and sends them to the corresponding edge terminal devices.

[0078] Each strategy model structure consists of the following: The recommendation weight correction parameter is used to correct the recommendation weight output by collaborative filtering in the S320 stage; The weight configuration and activation rule combination of the policy rule plug-in is used to replace or update the existing policy plug-in in the edge terminal; User group portrait mapping rules are used to map user behaviors into labels and correspond to strategy adaptation logic in subsequent recommendations.

[0079] After receiving the update policy parameters, the edge terminal performs the following operations: Load the updated recommendation weight correction logic and replace the original weight combination function; Hot update policy rule plug-in module, activate the new policy configuration, and close the old version plug-in; Bind the policy version with the current timestamp and device ID to generate a record of the current policy instance for subsequent data collection and feedback paths.

[0080] The updated model parameters will be applied to the S320 recommendation weight correction and S330 recommendation result sorting in the next round of recommendation, realizing dynamic strategy adaptation, improved recommendation accuracy and closed recommendation feedback loop.

[0081] This step implements the recommendation system's strategy self-evolution mechanism by designing a complete closed loop of user feedback collection and strategy training. The local encryption and compression of user click behavior and redemption results improves the security and efficiency of data transmission, and the federated learning mechanism ensures strategy optimization capabilities based on data privacy. The updated strategy parameters ultimately output by the system are sent to the edge terminal, effectively improving the timeliness, personalization, and dynamic response capabilities of the recommendation list generation strategy, and providing the latest strategy support foundation for subsequent meta-learning modules and lifecycle adjustment modules, thereby building an intelligent recommendation closed-loop system that drives strategy evolution with behavioral feedback.

[0082] Step 500 at least includes steps S510-S530: S510: Obtain unclicked equity items in the real-time recommendation list data, and construct a negative sample set for feature annotation.

[0083] This step receives the real-time recommendation list data output by the S330 module, which contains the set of benefits items displayed during the user's current recommendation period:

[0084] In this set, the system identifies a subset of benefits for which the user did not click:

[0085] For each unclicked benefit , the system extracts its embedded features with the user , carton embedding features , location information embedding Construct the negative sample feature vector together:

[0086] Construct supervised learning data pairs:

[0087] in: The set of equity items pushed to the user by the system in the current recommendation period t; is the specific equity item i in the recommendation period t; For The set of unclicked equity items filtered out; The user has not clicked on the benefit item; is the graph neural network embedding vector for the corresponding user; is the embedding vector of the cartons bound to the user in the current redemption scenario; The embedding vector of the user's location when scanning the code; Input features for the concatenated negative samples; is a supervisory label, representing “not accepted”; S520: Based on the negative sample set and the update strategy parameters, train the meta-learning model to generate new equity conversion rules.

[0088] In order to improve the model's generalization ability for complex non-click behaviors, functional analysis and variational methods are introduced to construct optimization objectives, and guided training is performed based on updating strategy parameters.

[0089] ① Construction model of equity substitution function: Introducing a class of strategy mapping functions , the input features Mapped to equity strategy weight vector:

[0090] in: is the strategy recommendation function, which is composed of parameters Control, from the feature space Mapping to policy space ; For the The composite feature vector of unclicked equity; The first The equity recommendation strategy vector corresponding to the sample.

[0091] ② Design an optimal path variation model to evaluate strategic objectives:

[0092] in: For training Meta-learning loss function of ; Indicates the number of negative samples currently used to train the meta-learning model; is the policy function parameter of the current training; It is a candidate replacement equity item; Candidate replacement equity Feature embedding of is a set of alternative equity items; is the square of the Euclidean distance; is the weight coefficient (e.g. it can be set to 0.2, the empirical value comes from historical iterations); A measure of the structural differences between the candidate equity item and the original equity item, which may include category, face value, applicable scenario, etc. Use target: Optimized This is the functional expression of the latest equity conversion rules.

[0093] S530: Encapsulate the new equity conversion rule into a structured plug-in, and load it into the policy rule plug-in module for subsequent recommendation and call.

[0094] The final trained function Converted into a structured strategy plug-in with the following structure:

[0095] in: Encapsulates the plug-in structure, including ID, policy function, equity mapping table and priority score; It is the unique identifier of the policy plugin, used to identify the current version; is the final equity mapping function; A mapping table between the original equity item and the recommended alternative item output by the policy function; It is the priority parameter of the policy plug-in.

[0096] This module introduces functional optimization and variational inference to model unclicked equity samples, combined with federated learning strategy parameter updates to design a generalizable equity replacement mapping model. The resulting plug-in strategy structure is not only highly interpretable but also supports pluggable deployment, enabling the system to self-evolve and self-adjust. This effectively improves the response efficiency, coverage, and user satisfaction of equity recommendations, demonstrating the system's practicality in real-world scenarios and the comprehensive improvement of its intelligent recommendation capabilities.

[0097] Step S600 at least includes steps S610-S630: S610: Obtain the carton scanning frequency and position migration data reported by the scanning terminal, and build a carton flow trajectory record.

[0098] Specifically, the scanning terminal will synchronously report the following data items each time a user scans a code: carton identification code, scanning timestamp, scanning location, terminal identifier, and operation type. This data is initially filtered and normalized by the multimodal data acquisition module. It is then matched and aligned with the user identification and location information in the scanning behavior data to form a time-sequenced carton operation sequence.

[0099] In the sequence structure, each record contains a unique carton code, user ID, latitude and longitude coordinates, geographic tags, scanning time and environmental context tags in a unified format. This data will be transmitted to the lifecycle identification module as the basis for trajectory construction.

[0100] Furthermore, the system constructs a directed temporal graph structure centered around the carton. Nodes are user-location associations, and edge weights are functions of time increments and migration distances. Each edge in the graph represents a carton's actual spatial location change. This graph structure allows for the continuous restoration of the carton's lifecycle flow links and their stability distribution.

[0101] S620: Estimate the remaining use period of the carton according to the carton flow trajectory record, and generate a carton life cycle label.

[0102] After the trajectory graph structure is generated, the system quantitatively evaluates the current life cycle status of the carton and performs comprehensive modeling based on the following indicators: Cumulative scanning frequency: the total number of times the carton is scanned within a unit of time; Number of cross-city migrations: The geographical span is determined by the number of changes in geographical tags; Average moving interval: Count the average time intervals between consecutive scans; User diversity coefficient: quantifies the degree of dispersion of the user groups using the cardboard box; Time difference since the first scan: estimate the length of time the carton has been used; Cumulative moving distance: Calculate the total length of actual displacement based on the trajectory points.

[0103] The above indicators are weighted and normalized to form a synthetic life cycle assessment function, and the carton life cycle labels are generated with reference to the prior model. The labels are divided into four categories: "initial period", "active period", "decline period" and "disposal period".

[0104] The system also jointly models the lifecycle label with data such as usage activity and user behavior distribution in the current period to form a lifecycle label vector with state-awareness. This label vector will serve as the core adjustment parameter for subsequent recommendation strategy priority adjustments.

[0105] S630: Combining the carton lifecycle tag with the policy rule plug-in, adjusting the priority of the benefit type in the real-time recommendation list data to generate updated policy data.

[0106] After obtaining the lifecycle tag, the system first reads the "Benefit Type and Lifecycle Mapping Rule Table" stored in the policy rule plug-in module. This table defines the set of benefit types and priority weight distribution that are applicable to each lifecycle stage.

[0107] The system matches the benefit item type in the real-time recommendation list with the current carton lifecycle tag and pushes priority updates based on the following logic: If the carton lifecycle tag is "active," prioritize user engagement benefits (such as points redemption and logistics tracking). If the life cycle is "disposable", the priority of environmental recycling benefits (such as green cashback and points incentives) will be increased; If the life cycle is in the "initial stage", the weight of incentive benefits (such as first-time code scanning coupons and behavior guidance benefits) will be increased.

[0108] The specific priority adjustment is performed by the push weight adjustment function in the policy rule plugin. This function accepts the lifecycle label vector and the set of equity item IDs in the current recommendation list as input and outputs an updated equity type weight matrix. Finally, the recommendation generation module reorders the recommended items based on this matrix and generates updated recommendation strategy data for the next step of recommendation feedback and strategy learning.

[0109] To form a logical closed loop, the updated strategy data will also be input as a feedback parameter into the recommendation weight correction process in the next round S310 to ensure dynamic optimization and lifecycle sensitivity of the recommendation process.

[0110] Through the design and operation of the above-mentioned S600 module, the system realizes the refined identification of the life cycle status of the carton, and dynamically adjusts the priority of the benefit push strategy based on the life cycle label, effectively enhancing the contextual adaptability and life cycle sensitivity of the recommended content, thereby improving the timeliness, relevance and user response efficiency of the benefit recommendation, and building a carton benefit intelligent recommendation system that combines intelligent decision-making and sustainable circulation.

[0111] Example 2: Figure 2 The structural block diagram of the carton intelligent rights exchange recommendation system based on multimodal data fusion and dynamic portrait according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include: Multimodal data acquisition module 10 is used to obtain the physical attribute data of the carton, the user's scanning behavior data and location information, and perform synchronous filtering and structured coding processing on the collected data to generate standardized feature data. This module includes: Receive physical parameters such as carton type, size, and material through the terminal device scanning code; Collect user scanning behavior data such as scanning time, frequency, and time period distribution; Extract spatial information such as the user's current location and trajectory movement; The above three types of data sources are format checked, missing data are completed, timestamps are aligned, and multimodal normalization is performed, and uniformly encoded into a structured standardized feature data stream for subsequent module calls.

[0112] Heterogeneous graph construction and embedding calculation module 20 is used to construct a heterogeneous graph structure consisting of user nodes, carton nodes, and equity nodes based on standardized feature data, perform connection strength analysis and edge pruning operations, and execute the graph neural network model on the pruned graph structure to generate user embedding data and carton embedding data. This module includes: Heterogeneous node mapping and edge weight initialization; Build a node association graph based on the frequency of code scanning and user historical behavior; Dynamically update graph connection weights and prune weakly connected edges; Utilize the graph neural network model to perform multiple rounds of node representation propagation and extract high-dimensional embedding features.

[0113] The recommendation generation and policy injection module 30 is used to perform collaborative filtering similarity calculation based on user embedded data and carton embedded data, combined with the user's historical redemption records, and input the results into the policy rule plug-in module to perform recommendation weight correction and ultimately generate real-time recommendation list data. This module includes: Comparison of high-dimensional features between similar users and similar cartons; Extraction of historical behavior influencing factors and preference trends; Policy rule plug-ins regulate logic loading, including equity priority weights, blacklist filtering, and regional restrictions; Output a list of highly matching interests that meet the policy conditions.

[0114] User behavior collection and federated training module 40 is used to collect user click behavior and redemption results on real-time recommendation list data, encrypt and compress the data locally, and then upload it to the cloud server. It aggregates data from multiple edge terminals, performs federated learning training to generate updated policy parameters, and transmits the updated parameters back to the edge terminal to replace the original model parameters. This module includes: Encrypted collection of local user interaction events; Behavioral feature compression and encoding; Federation server model aggregation mechanism; Fine-tune policy parameters and implement distribution updates.

[0115] Rule generation and plug-in packaging module 50 is used to obtain unclicked equity items in the real-time recommendation list data, construct a negative sample set, and generate new equity conversion rules by training the meta-learning model based on updated policy parameters. The generated rules are packaged as a structured plug-in and loaded into the policy rule plug-in module. This module includes: Build a negative sample library and perform multi-dimensional feature annotation; Modeling of benefit type, rejection frequency, and user grouping; Adjust rule generation weights based on a meta-learning framework; Output hot-swappable policy plug-ins, supporting rule updates and dynamic loading.

[0116] Lifecycle Identification and Policy Adjustment Module 60 is used to obtain the carton scanning frequency and location migration data reported by the scanning terminal, construct a carton flow trajectory record, and estimate its lifecycle label based on the trajectory. In conjunction with the policy rule plug-in module, it adjusts the priority of the equity type in the real-time recommendation list data to generate updated policy data. This module includes: Construct a time series matrix of carton usage behavior sequences and trajectories; Identification of life cycle stages, such as "initial use period", "active period", "idle period", and "near waste period"; Call the lifecycle policy template to adjust the type and distribution weight of recommended benefits; Update the recommendation strategy configuration table and push it to the recommendation generation module to achieve closed-loop optimization.

[0117] This embodiment builds a closed-loop system covering the entire process from original behavior perception, graph modeling, personalized recommendation, feedback training, rule evolution to lifecycle recognition, achieving the following beneficial effects: (1) Enhanced multimodal data fusion capabilities: By integrating the physical properties of cartons, user scanning behavior, and location information, standardized data semantics processing is achieved, which improves the perception ability and scenario adaptability of the recommendation system.

[0118] (2) Improve the accuracy and real-time performance of recommendations: The graph neural network embedding features combined with collaborative filtering and strategy plug-in optimization significantly improve the personalization level and timeliness of equity matching.

[0119] (3) Implementing federated learning-based secure training: Adopting local encrypted data collection and cloud-based aggregated training mechanisms to protect user privacy while dynamically optimizing recommendation strategy parameters, and achieving cross-terminal learning adaptability.

[0120] (4) Supporting the linkage between strategy plug-ins and life cycle: The negative sample-driven meta-learning model supports the autonomous evolution of strategy plug-ins. Combined with the carton life cycle management, it effectively guides user behavior, extends the carton circulation benefit cycle, and improves the intelligent decision-making ability of the overall system.

[0121] (5) Forming a closed-loop self-optimization capability chain: forming a complete closed loop from data input to result feedback, with the ability of continuous evolution, autonomous learning and rule iteration, providing a stable and efficient intelligent recommendation basis for equity exchange.

[0122] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. A carton intelligent rights exchange recommendation method based on multimodal data fusion and dynamic profiling, characterized by: The following steps are involved: Obtain the physical attribute data of the carton, the user's scanning behavior data and location information, perform synchronous filtering and structured coding, and generate standardized feature data; Based on the standardized feature data, a heterogeneous graph structure of user nodes, carton nodes, and equity nodes is constructed, connection strength analysis and pruning are performed, and user embedding data and carton embedding data are generated through a graph neural network; Based on the user embedded data and carton embedded data, combined with the user's historical redemption records, collaborative filtering recommendations are performed, and a strategy rule plug-in is loaded to modify the recommendation weights to generate real-time recommendation list data; Collecting the user's click behavior and redemption results on the real-time recommendation list data, encrypting and compressing it locally, and uploading it to the cloud to perform federated learning training and generate update strategy parameters; Obtain unclicked equity items in the real-time recommendation list data, construct a negative sample set, train a meta-learning model based on the updated policy parameters, generate equity conversion rules, and encapsulate them as a plug-in to be loaded into the policy rule plug-in module; The updating strategy parameter training meta-learning model includes: ； in, For training Meta-learning loss function of ; Recommend function for strategy; Indicates the number of negative samples currently used to train the meta-learning model; is the policy function parameter of the current training; For the The composite feature vector of unclicked equity; It is a candidate replacement equity item; is a set of alternative equity items; Candidate replacement equity Feature embedding of is the square of the Euclidean distance; is the weight coefficient; It measures the structural differences between candidate equity items and original equity items; obtains the carton scanning frequency and position migration data reported by the scanning terminal, constructs a carton flow trajectory record, generates a carton life cycle label, and adjusts the equity type priority in the real-time recommendation list data in combination with the policy rule plug-in to generate update policy data.

2. The method according to claim 1, characterized in that Generating standardized feature data comprises the following steps: Obtain the physical attribute data of the carton, user scanning behavior data and location information; Perform feature extraction and normalization on the physical attribute data of the carton to generate physical attribute features; The user's scanning behavior data and location information are time-series encoded and fused with the physical attribute features to generate standardized feature data.

3. The method according to claim 1, characterized in that Generating user embedding data and carton embedding data comprises the following steps: Constructing a heterogeneous graph structure including user nodes, carton nodes, and equity nodes based on the standardized feature data; Performing connection strength analysis on the heterogeneous graph structure and performing edge pruning processing; Graph neural network embedding processing is performed on the pruned heterogeneous graph structure to generate user embedding data and carton embedding data.

4. The method according to claim 1, wherein Generating real-time recommendation list data includes the following steps: Based on the user embedding data and the carton embedding data, combined with the user's historical redemption records, calculating collaborative filtering similarity; The similarity calculation results are input into the policy rule plug-in module, and the rule weights are loaded to perform the recommended weight correction; A preset number of equity items are filtered according to the revised recommendation weights to generate real-time recommendation list data.

5. The method according to claim 1, wherein Collecting user behavior and performing federated learning training includes the following steps: Collecting the user's click behavior and redemption results on the real-time recommendation list data, and performing encryption and structural compression; Upload encrypted and compressed user click behaviors and redemption results to the cloud, aggregate multi-terminal data, and perform federated learning training; Generate update policy parameters and prepare to send them to edge terminals.

6. The method according to claim 1, characterized in that The following steps are also included: Sending the update policy parameters to the edge terminal; Replace the original model parameters at the edge terminal; Participate in subsequent recommendation weight correction operations based on the updated model parameters.

7. The method according to claim 1, characterized in that Generating the equity conversion rules includes the following steps: Obtaining unclicked equity items in the real-time recommendation list data to construct a negative sample set; A meta-learning model is trained based on the negative sample set and the updated strategy parameters to generate new equity conversion rules; The equity conversion rules are encapsulated as a structured plug-in and loaded into the policy rule plug-in module.

8. The method according to claim 1, characterized in that Generating a carton life cycle label comprises the following steps: Obtain carton scanning frequency and location migration data reported by scanning terminals to build a carton flow trajectory record; Estimate the remaining use cycle of the carton based on its circulation trajectory and generate a carton life cycle label; Output the carton life cycle label for subsequent recommendation strategy use.

9. The method according to claim 1, characterized in that Generating update policy data comprises the following steps: Combining the carton lifecycle tag with the policy rule plug-in; Adjusting the priority of the equity type in the real-time recommendation list data; Generate updated recommendation strategy data.

10. A carton intelligent rights exchange recommendation system based on multimodal data fusion and dynamic portrait, applied to the carton intelligent rights exchange recommendation method based on multimodal data fusion and dynamic portrait according to any one of claims 1 to 9, characterized in that: include: Multimodal data acquisition module, used to obtain the physical attribute data of the carton, user scanning behavior data and location information, and generate standardized feature data; A heterogeneous graph construction module, configured to construct a heterogeneous graph structure based on the standardized feature data and perform embedding calculations to generate user embedding data and carton embedding data; A recommendation generation module, configured to perform collaborative filtering calculations based on the embedded data and user history records, and to call a policy rule plug-in to generate real-time recommendation list data; Feedback training module, which is used to collect user feedback, perform federated learning training, and issue updated policy parameters; The rule generation module is used to construct a negative sample set, train the meta-learning model to generate equity conversion rules, and encapsulate them as a plug-in and load them into the policy rule plug-in module; The life cycle identification module is used to generate carton life cycle labels based on carton flow trajectory data, adjust the recommendation strategy priority, and generate update strategy data.

Images